Update README.md

README.md

@@ -1,69 +1,168 @@
 ---
 base_model: unsloth/Qwen3-0.6B-unsloth-bnb-4bit
+model_name: Math Mini 0.6B (Preview)
 tags:
 - text-generation
 - text-generation-inference
 - transformers
-- unsloth
 - qwen3
-- gguf
 - math
 - enosis-labs
 - math-mini
+- gguf # Tag for GGUF version
 license: apache-2.0
 language:
 - en
-library_name: llama.cpp # O la principal librería para GGUF que recomiendes
-pipeline_tag: text-generation
 ---
 
-# Math Mini 0.6B (Preview) - GGUF
+# Math Mini 0.6B (Preview)
 
-## Model Card for `enosislabs/math-mini-0.6B-preview-gguf`
+**Math Mini 0.6B (Preview)** is a compact, specialized model developed by **Enosis Labs** as part of the "Mini Series." It is designed to deliver efficient and precise mathematical reasoning, with a realistic and practical focus for its size. This model is fine-tuned from `unsloth/Qwen3-0.6B-unsloth-bnb-4bit`.
 
-<p align="center">
-  <img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="170"/>
-</p>
+## Philosophy & Capabilities
 
-**Math Mini 0.6B (Preview)** is a specialized language model developed by **Enosis Labs**, fine-tuned for mathematical problem-solving. This compact 0.6 billion parameter model is part of our "Math Mini" series, designed for efficiency and accessibility, particularly for performing mathematical reasoning tasks.
+The Mini Series, along with the "Enosis Math" and "Enosis Code" models, incorporates step-by-step reasoning by default, enabling more efficient, clear, and well-founded answers. All models in the Math series have been trained with carefully curated step-by-step problem-solving datasets, resulting in a greater ability to reason and explain solutions in a structured way.
 
-This GGUF version is quantized and optimized for running locally on CPUs and other supported hardware through interfaces like `llama.cpp`, `LM Studio`, `Oobabooga Text Generation WebUI`, etc.
+**Math Mini 0.6B (Preview)** is optimized for:
 
-**Please note: This is a preview version.** We are actively working on improvements and further evaluations.
+* **Basic Algebra:** Solving equations and manipulating expressions.
+* **Arithmetic & Sequential Reasoning:** Calculations and breaking down problems into logical steps.
+* **Elementary Logic:** Applying deduction in mathematical contexts.
+* **Introductory Competition Problem Solving:** Focus on foundational skills adapted to the model's scale.
 
-## Model Details
+Larger models in the "Enosis Math" series address advanced topics such as calculus, higher algebra, and olympiad problems. The "Code Mini" and "Enosis Code" series are oriented towards programming and algorithmic tasks, maintaining the same philosophy of explicit and efficient reasoning.
 
-* **Developed by:** [Enosis Labs](https://huggingface.co/enosislabs) (¡Asegúrate de que este enlace sea correcto o créalo!)
-* **Model type:** Qwen3-based Decoder-only Transformer
-* **Series:** Math Mini
-* **Version:** 0.6B (Preview)
-* **Language(s):** Primarily English (enfocado en problemas matemáticos en inglés)
-* **License:** Apache 2.0
-* **Finetuned from:** `unsloth/Qwen3-0.6B-unsloth-bnb-4bit`
-* **Fine-tuning Frameworks:** [Unsloth](https://github.com/unslothai/unsloth) for 2x faster training and optimized memory usage, in conjunction with Hugging Face's [TRL](https://github.com/huggingface/trl) library.
-* **Fine-tuning Data:** Fine-tuned on a curated collection of mathematical reasoning datasets, including algebraic problems, multi-step arithmetic, and word problems. Key datasets included variants of AQuA-RAT and OpenMathReasoning.
+This model is a **preview version** and is under continuous improvement and evaluation.
 
-## Intended Uses
+## Quick Start
 
-This model is intended for (but not limited to):
+Available in both Hugging Face Transformers and quantized GGUF formats.
 
-* Assisting with grade-school to early high-school level mathematical problems.
-* Solving algebraic equations and multi-step arithmetic.
-* Understanding and processing mathematical questions posed in natural language.
-* Use as a lightweight math reasoning engine in applications where resource efficiency is crucial.
-* Educational purposes, experimentation, and as a foundational block for further specialization.
+### Transformers (Hugging Face)
 
-## How to Get Started with GGUF
+Ensure you have the latest `transformers` library. For Qwen3 models, a recent version is recommended.
 
-This model is provided in GGUF format, making it easy to run locally.
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
 
-**1. Download the GGUF file:**
-   You can find the GGUF file(s) in the "[Files and versions](https_url_del_repositorio_en_hf/tree/main)" tab of this repository (reemplaza `https_url_del_repositorio_en_hf` con el enlace real a tu repo). Look for files ending with `.gguf`. For example: `math-mini-0.6b-preview-q4_K_M.gguf` (o el nombre exacto de tu archivo).
+model_id = "enosislabs/math-mini-0.6b-preview-gguf"
 
-**2. Using with `llama.cpp` (Example):**
+pipe = pipeline("text-generation", model=model_id, trust_remote_code=True)
 
-   First, ensure you have `llama.cpp` compiled.
-   ```bash
-   git clone [https://github.com/ggerganov/llama.cpp.git](https://github.com/ggerganov/llama.cpp.git)
-   cd llama.cpp
-   make
+messages = [
+    {"role": "system", "content": "You are a helpful math assistant."},
+    {"role": "user", "content": "Solve for x: 3x + 11 = 35"},
+]
+
+formatted_prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+outputs = pipe(formatted_prompt, max_new_tokens=100)
+print(outputs[0]["generated_text"])
+
+# Alternatively, load the model and tokenizer directly:
+tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
+inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")
+outputs = model.generate(inputs, max_new_tokens=100)
+response_text = tokenizer.decode(outputs[0], skip_special_tokens=False)
+print(response_text)
+```
+
+### GGUF with Ollama
+
+Download the `.gguf` file from Hugging Face and use Ollama. You can choose between other GGUF versions such as 4bit, 5bit, and 8bit. This example uses the 4bit version:
+
+```bash
+ollama run enosislabs/math-mini-0.6b-preview-gguf:Q4_K_M
+```
+
+For more control, create a `Modelfile` with the Qwen3 template:
+
+```modelfile
+FROM ./math-mini-0.6b-preview-Q4_K_M.gguf
+TEMPLATE """
+<|im_start|>system
+{{ .System }}<|im_end|>
+<|im_start|>user
+{{ .Prompt }}<|im_end|>
+<|im_start|>assistant
+"""
+```
+
+Then run:
+
+```bash
+ollama create math-mini-0.6b -f Modelfile
+ollama run math-mini-0.6b
+```
+
+### GGUF with llama.cpp
+
+```bash
+./main -m ./path/to/math-mini-0.6b-preview.gguf -n 256 -p "<|im_start|>system\nYou are a helpful math assistant.<|im_end|>\n<|im_start|>user\nSolve for x: 2x + 5 = 15<|im_end|>\n<|im_start|>assistant\n" --temp 0.2 -c 2048
+```
+
+### vLLM (Transformers)
+
+```bash
+pip install vllm
+python -m vllm.entrypoints.openai.api_server --model enosislabs/math-mini-0.6b-preview-gguf --trust-remote-code
+```
+
+For chat:
+
+```bash
+curl -X POST "http://localhost:8000/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    --data '{
+        "model": "enosislabs/math-mini-0.6b-preview-gguf",
+        "messages": [
+            {"role": "system", "content": "You are a helpful math assistant."},
+            {"role": "user", "content": "What is the capital of France?"}
+        ],
+        "max_tokens": 50,
+        "temperature": 0.2
+    }'
+```
+
+## Prompt Format (Qwen3 ChatML)
+
+For best results, use the Qwen3 ChatML format. The `tokenizer.apply_chat_template` method handles this automatically.
+
+```text
+<|im_start|>system
+You are a helpful AI assistant. Provide a detailed step-by-step solution.
+<|im_end|>
+<|im_start|>user
+{user_question}
+<|im_end|>
+<|im_start|>assistant
+```
+
+## Acknowledgements
+
+* Fine-tuned from `unsloth/Qwen3-0.6B-unsloth-bnb-4bit`.
+* Training process accelerated and optimized thanks to [Unsloth](https://github.com/unslothai/unsloth).
+
+## Citation
+
+If you use this model, please cite:
+
+```bibtex
+@software{enosislabs_math_mini_0.6b_preview_2025,
+  author = {{Enosis Labs}},
+  title = {{Math Mini 0.6B (Preview)}},
+  year = {2025},
+  publisher = {Hugging Face},
+  version = {0.1-preview},
+  url = {https://huggingface.co/enosislabs/math-mini-0.6b-preview-gguf}
+}
+```
+
+<!--
+Key points:
+- More subtle and direct, less redundancy.
+- Emphasizes default activation of step-by-step reasoning across the series.
+- Clear and modern examples for each format.
+- ChatML prompt is central to the experience.
+- Assumes the repo contains both Transformers and GGUF models.
+-->